Deep Learning-Assisted Spoken-Command Classification Using HVP/CNT Sponge Pressure Sensors

Abstract

Flexible piezoresistive pressure sensors capable of resolving weak dynamic mechanical signals are important for wearable health monitoring, human-machine interaction, and vibration-pattern recognition. However, integrating high sensitivity, a broad working range, cyclic repeatability, and machine-learning-based signal discrimination in a lightweight porous device remains challenging. Herein, an ultralight HVP/CNT sponge pressure sensor (apparent density 0.08–0.15 g cm–3, porosity 88–93%) was fabricated by incorporating hydroxylated multiwalled carbon nanotubes (MWCNT–OH) into a porous hydrolysis-stable HVP resin framework. The optimized HVP/CNT sensor showed a peak sensitivity of 9.17 kPa–1 in the low-pressure region, distinguishable piezoresistive responses over the tested pressure range, response/recovery times of 360/470 ms, and broadly maintained output during 2500 loading–unloading cycles under the tested conditions. The sensor monitored representative human-motion signals and captured throat-vibration signals during phonation. After the relative resistance-change signals were converted into image-like features and processed by an 11-layer convolutional neural network (CNN), same-participant classification accuracies of 97.625% for two language categories and 97.25% for eight spoken commands were obtained under controlled acquisition conditions. These results demonstrate a materials-device-algorithm route for wearable speech-related vibration classification; extension to industrial condition monitoring should be supported by direct equipment-vibration tests, environmental cross-sensitivity evaluation, and longer-term cyclic durability analysis.